Position-dependent phonetic models for reliable pronunciation identification

• US 20080172224A1
• Filed: 01/11/2007
• Published: 07/17/2008
• Est. Priority Date: 01/11/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Shi, Yu, Liu, Peng, Soong, Frank Kao-ping
• Granted Patent: US 8,135,590 B2
• Time in Patent Office: Days
• US Class Current: 704/220
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

A representation of a speech signal is received and is decoded to identify a sequence of position-dependent phonetic tokens wherein each token comprises a phone and a position indicator that indicates the position of the phone within a syllable.

Claims

1. A method comprising:

receiving a representation of a speech signal; and

decoding the representation of the speech signal to identify a sequence of position-dependent phonetic tokens wherein each token comprises a phone and a position indicator that indicates the position of the phone within a syllable.

2. The method of claim 1 further comprising annotating a lexicon with the sequence of position-dependent phonetic tokens.

3. The method of claim 1 wherein decoding the representation of the speech signal comprises using a word language model to decode the representation into a sequence of words and converting the sequence of words into the sequence of position-dependent phonetic tokens.

4. The method of claim 3 wherein converting the sequences of words into the sequence of position-dependent phonetic tokens comprises retrieving a position-dependent phonetic token description of each word from a lexicon.

5. The method of claim 3 further comprising applying the sequence of position-dependent phonetic tokens to a position-dependent phonetic token language model that provides probabilities for sequences of position-dependent phonetic tokens to determine at least one probability for the sequence of position-dependent phonetic tokens.

6. The method of claim 5 further comprising utilizing the at least one probability as a confidence measure.

7. The method of claim 1 wherein each position indicator indicates one of a group of syllable positions consisting of start phone, end phone, onset consonant, coda consonant, middle vowel, and single vowel.

8. A computer-readable medium comprising computer-executable instructions for performing steps comprising:

decoding a representation of a speech signal using a position-dependent phonetic language model that provides probabilities of sequences of position-dependent phonetic tokens wherein each position-dependent phonetic token comprises a phone and a position identifier that identifies a position within a syllable, wherein decoding produces at least one sequence of position-dependent phonetic tokens.

9. The computer-readable medium of claim 8 having computer-executable instructions for performing further steps of storing one sequence of position-dependent phonetic tokens as a pronunciation for a sequence of symbols in a lexicon.

10. The computer-readable medium of claim 9 wherein the sequence of symbols is not a word in a language.

11. The computer-readable medium of claim 8 wherein decoding produces a true-transcript sequence of position-dependent phonetic tokens representing a true-transcript of the speech signal and at least one alternative sequence of position-dependent phonetic tokens.

12. The computer-readable medium of claim 11 having computer-executable instructions for performing further steps comprising using the true-transcript sequence of position-dependent phonetic tokens and the at least one alternative sequence of position-dependent phonetic tokens to discriminatively train an acoustic model.

13. The computer-readable medium of claim 8 wherein a position identifier is one of a group of position identifiers consisting of start phone, end phone, onset consonant, coda consonant, middle vowel, and single vowel.

14. A method comprising:

identifying a position-dependent phonetic description of each word in a word language model, each position-dependent phonetic description comprising at least one position-dependent phonetic token, each position-dependent phonetic token comprising a phone and a position identifier that identifies a position within a syllable;

using probabilities for words and probabilities for sequences of words in the word language model to set probabilities for position-dependent phonetic tokens and probabilities for sequences of position-dependent tokens to construct a position-dependent phonetic language model; and

storing the position-dependent phonetic language model.

15. The method of claim 14 wherein setting probabilities comprises developing within-word counts for position-dependent phonetic tokens and for sequences of position-dependent phonetic tokens that are bounded within individual words based on unigram probabilities of the words that the position-dependent phonetic tokens are bound within.

16. The method of claim 15 wherein setting probabilities comprises developing cross-word counts for sequences of position-dependent phonetic tokens that span multiple words based on probabilities of the sequences of multiple words that the sequences of position-dependent phonetic tokens span.

17. The method of claim 16 wherein setting probabilities further comprises using the within-word counts and the cross-word counts to determine a joint probability for position-dependent phonetic tokens in a sequence of position-dependent phonetic tokens.

18. The method of claim 14 further comprising using the position-dependent phonetic language model to identify a sequence of position-dependent phonetic tokens from a speech signal.

19. The method of claim 18 further comprising using the sequence to annotate a lexicon with a pronunciation for at least one symbol that is not found in a written language.

20. The method of claim 18 further comprising using the sequence of position-dependent phonetic tokens as part of discriminative training of an acoustic model